Airline Business Models and Their Network Structures Anne Lange, Tobias Bier

Logistics Research (2019) 12:6 DOI_10.23773/2019_6

Airline business models and their network structures Anne Lange, Tobias Bier

Received: 01 June 2018 / Accepted: 27 May 2019 / Published online: 24 June 2019 © The Author(s) 2018 This article is published with Open Access at www.bvl.de/lore

ABSTRACT 1 INTRODUCTION

Network structure is one parameter that feeds into Airlines provide air transportation service to airline business models. We analyze the networks of passengers and cargo. To do so, they operate a fleet of 58 European airlines including full service carriers, aircraft and make use of them to transport passengers low cost carriers, regional and charter airlines. Eight and cargo from one place to another. The underlying network metrics are calculated to describe the various route network is essential for the operations. The route aspects of network structure. A principal component network is the entirety of flights offered by an airline. analysis is conducted and indicates two components in Hence, it is the backbone of the operations. Given that the metrics. The components address network coverage aircraft and fuel expenses constitute major costs for and service network. The airline networks are clustered airlines, route network decisions are tightly linked to based on the identified components. Findings from the the cost of operations for an airline. At the same time, analysis include that the resulting clusters based on the destinations, connecting services, flight frequencies only network structure appear to be consistent groups and travel time offered by an airline define the essential of airline business models. It indicates that only judging product offered by the airline to its customers. from network structure allows to reason on airline Airline business models explain the underlying business models. In more detail, full service carriers rationale on how an airline provides service to are structured in three subgroups differing in coverage its customers and airline network structure is an as well as in the operated services. At the same time, established element of airline business models. One low cost carriers, charter and regional airlines appear common assumption is that Full Service Carriers as clusters in the analysis. A few airlines are identified (FSC) tend to operate hub and spoke networks (HS) as outliers and investigating their business model whereas and Low Cost Carriers (LCC) operate point- confirms this network observation. to-point networks (PP). The structure of transportation networks has been KEYWORDS: metrics · graph theory · principal studied in recent years and advancements have been component analysis · airline business model · network made on describing various aspects of transportation structure network structure quantitatively. For instance, Hu and Zhu [1] describe maritime networks, Wang and ‘‘This article is part of a focus collection on ‘‘Understanding Cullinane [2] work on seaport centrality, Dobruszkes Future Logistics’’ based on the BVL’s 9th International Scientific [3] illustrates LCC network structures, and Mishra et al. Symposium on Logistics in Magdeburg in 2018“ [4] suggest metrics for the connectivity of urban public transport networks. Mattsson and Jenelius [5] review Anne Lange research on transportation network vulnerability, University of Luxembourg, Luxembourg Centre one of the typical application areas where network for Logistics and Supply Chain Management, structures need to be quantified. 162A, avenue de la Faïencerie, L-1511 Luxembourg, Luxembourg, Thus far, these two streams of research are not [email protected], yet well connected. That is, on the one hand there Tel: +352 46 66 44 6308 are extensive metrics to account for transportation network structure. On the other hand, airline networks Tobias Bier TU Darmstadt, are described on a rather high level when it comes to Department of Law and Economics, studying airline business models. Our work is positioned Hochschulstr. 1, at this interface. We aim to argue that a more detailed 64289 Darmstadt, Germany 2 operationalization of airline network structure benefits is relevant. Airline network structures have been the description of airline business models. Thus, we quantified by graph theoretic metrics commonly explore the research question in how far a quantitative used in social network analysis. The seminal work by assessment of airline network structures supports the Guimerá and Amaral [14] analyzes the worldwide air clear description of airline business models. We focus transportation network and showed that it belongs to solely on exploring the information contained in the the classes of scale-free and small world networks, airline network structure. In doing so, we contribute indicating an efficient connectivity. To do so, the authors findings to airline business model research regarding study (scaled) node degree and node betweenness the description of airline networks and wish to support centrality. Similar approaches have been applied to the development of more clear-cut models of airline national networks, for instance by Guida and Maria business models. [15] for the Italian network (node degree and node Our approach is novel for airline business model betweenness) and by Wang et al. [16] for the Chinese analysis: We calculate established metrics of airline network (average path length, clustering coefficient, network structure and apply a principal component degree distribution, node degree, node closeness and analysis in order to reduce the dimensions of the dataset. node betweenness). Alderighi et al. [17] provide a set of This allows us to assess airline network structure from metrics to study the structure of airline networks. This multiple perspectives and account for the correlation set includes the Gini index and network betweenness of the different metrics. When clustering the airlines for spatial analysis. In their literature review, Lordan according to the network structures alone, similarities et al. [18] highlight that the analysis of airline network between airlines already appear that relate well to structure has received limited attention from airline similarities in airline business. management research. Instead, research has rather The structure of the remainder is as follows: Section been conducted mainly from the complex network 2 provides the necessary theoretical background on point of view. Some examples of studies in airline transportation network structure and airline business management research do exist, however. For instance, models. Section 3 summarizes the analyzed dataset of Reggiani et al. [19] analyze the network of Lufthansa 58 airlines and presents the metrics from graph theory as well as that of Star Alliance and convincingly used. It further outlines the principal component highlight the network impact of strategic choices made analysis we use to reduce the dimensionality in the by Lufthansa and Star Alliance. The authors make use collected data and shows how we apply a cluster analysis of a large set of metrics, including node degree, node to detect airlines of similar network structure. Section closeness, node betweenness, diameter, clustering 4 presents the interpretation of the factor loadings as coefficient, Gini index, network betweenness, and well as the findings from the cluster analysis. Section 5 entropy function. Bowen [20] describes the cargo discusses the findings and compares to prior research. networks of UPS and FedEx. He includes an analysis Section 6 concludes. of the network structures by studying the number of nodes, the number of edges, the beta and the gamma indices. Analyzing airline network structure is of 2. THEORETICAL BACKGROUND interest to various disciplines and certain graph theory metrics have repeatedly been used to do so. However, Choosing a suitable network structure is a complex yet no standard set of well-established metrics exists for important task for an airline and it is not surprising this purpose. Further, the multitude of metrics used to that it has received attention in the literature. Early account for the structure of airline networks oftentimes contributions address the issue from the perspective provides comparable yet not identical insights into the of changes in network structure as related to the structure. liberalization of aviation markets. The economic Since airline network structure is essential for the literature was interested in understanding the formation service and cost of airline operations, it is an element of HS in deregulated markets (see, for instance, Oum in airline business models. Passenger airline business et al. [6]). Pros and cons of HS as compared to PP models have traditionally been categorized into four have been studied (see, e.g., [7], [8]). More recent groups: full service carriers, low cost carriers, regional work results from the observation that FSC and LCC carriers and charter airlines. It is not trivial to clearly do, among others, compete based on their network distinguish between FSC and LCC since the business structure (e.g., [9]). Obviously, competition not only models have been evolving. Button and Ison [21] point results from the network structure. Babić and Kalić out that LCC, in general, implement specific strategies [10] argue that it is the interplay of network structure that differentiate them from FSC. Notably, LCC offer a and pricing policy that airlines compete on, whereas limited range of service in their basic fares as compared Brueckner [11] and Brueckner and Flores-Fillol [12] to FSC. That is, they strip the core service down to factor in average flight delay. the transportation and offer additional services at an There is a wide spectrum of potential airline network extra charge, creating additional revenues. Service is structures, extending far beyond the archetypes of HS focused on a single booking class and bookings are and PP [13]. Thus, describing them in more detail often possible only online. Airports served tend to Airline business models and their network structures 3 be second-tier airports. Further, LCC focus on short carriers are active only in a restricted geographic area. aircraft ground times and reduce cost by keeping the Charter airlines do not provide their service directly variation in aircraft types in the fleet low. Moreover, to passengers, but rather to wholesale intermediaries they make use of their market power towards suppliers such as travel and tourism companies. The operated to create efficiencies in sourcing [21]. It is common routes are requested by the intermediaries. However, to associate LCC with PP and FSC with HS. As such, as competition increases in the challenging airline LCC and FSC differ in both their service offer and their industry, the boundaries of airline business models are cost positions and compete against each other. Regional blurring and business models are evolving [22-25].

Table 1: Network aspects in airline business models

Reference Type Structural network aspects touched upon in identifying airline business models

Bachwich and Wittman [23] N Stage length

Casadesus-Masanell and Ricart [27] N Stage length

Daft and Albers [28] FQ Stage length, service frequency

Dobruszkes [3] FN Gamma connexity index, number of airports served, number of routes operated

Dobruszkes [29] FN Number of cities served, number of routes operated, centralization of the network

Gillen and Gados [25] N Stage length, geographic coverage, service frequency

Klophaus et al. [30] FQ Point-to-point service

Lohmann and Koo [31], Jean and FN Network density, number of destinations, Lohmann [32], Moir and Lohmann [33] service frequency, stage length

Mason and Morrison [34] FN Network density, number of routes, service frequency

O’Connell [35] N Hub-and-spoke operations

Pereira and Caetano [36] FQ Network architecture

Soyk et al. [24] FQ Network concentration, service frequency

Urban et al. [37] FQ Network system (Point-to-point, Hub-and-spoke, Multi-hub, Point-to-point & Hub-and-spoke), geographical coverage

Wensveen and Leick [26] FQ Stage length, number of destinations, frequency

Types: (FN) Framework mostly quantitative, (FQ) Framework mostly qualitative, (N) Narrative 4

Passenger airline business models describe and The dataset contains 134652 flights from 58 European explain how passenger airlines create and capture value. airlines during the week of November 19-26, 2017. Researchers do so by following various approaches. Airlines were selected based on their fleet size. In order Table 1 highlights, how the aspect of network figures in to draw conclusions regarding network structure, the the description of airline business models. It is striking airlines need substantial operations. Azur Air and TAP that network is typically considered at a simplified Express operate the smallest fleet in the sample with 22 level, taking into account one or very few features aircraft each. We include an airline if its headquarters of an airline network. Stage length appears to be the is located in a country on the European continent. most prominent feature included. Dobruszkes [3, 29] Hence, we include Russian and Turkish airlines in the addresses the airline network at a more granular level. sample. Table A.1 in the Appendix lists the airlines For one, he takes into account the gamma connexity included as well as their fleet size and country of origin. index as well as the centralization of the network, For each flight, data on the origin, destination and both of which relate to graph theory. Moreover, his operating date were collected. This information is used understanding of the operating network goes far to reconstruct the as-operated flight networks of the beyond its structure. In addition to the features in airlines. The networks are interpreted as graphs where Table 1, he includes parameters on how far the airline the nodes are the individual airports and directed edges network relies on the 5th through 9th freedom of the represent the connections of each airline between two air, in how far destinations tend to be ‘warm water’ airports in the timespan. The arcs are weighted by the destinations, the rivalry at the airports and the share number of flights on the connections per week. of international flights. In doing so, he consciously Note that the aircraft are allocated to the individual adds additional elements that describe the nature of airlines based on their registration. That is, wet leased the considered airline networks but that are not covered aircraft appear as flying for the owning company. by graph theory metrics. These elements are specific Further, code share flights are allocated to the operating to airline networks and provide the analyst with more airline. In addition, we do not aggregate airlines to their insights. mother companies, but allocate aircraft to airlines solely based on their registration. This implies, for instance, that HOP! appears as an individual airline 3. METHODOLOGY in our sample, independent from its mother company Air France. This seems reasonable as HOP! acts on the market independently from Air France at the time 3.1. Dataset of the data collection. It is certainly true that this We collected flight data to be analyzed from approach may be challenged as it also implies that TAP flightradar24.com. Our analysis is based on the Express appears individually even though it operates as information of flights that were operated by the a capacity provider to TAP airlines. The key question individual airlines during the observation period. The to be addressed would be to identify which flights are data we collected originates from Automatic Dependent actually planned under individual responsibility and Surveillance-Broadcast (ADS-B) technology. Its which ones are jointly managed with another airline. As benefit for our purposes is that it retraces aircraft as this question cannot be answered consistently without they move in the air and on ground, providing us with inside knowledge on all airlines, we refrained from this the information of operated flights. It can be used to approach apart from one exception. We aggregated observe and analyze flight trajectories (e.g. [38, 39]), Germanwings and Eurowings flights since it is public a feature that we do not even need to employ. We knowledge that at the time of data collection they had cross-checked a sample of the collected information already been merged and were operating jointly. with actual flight schedules and are confident that the information we collected accurately reflects flight operations. Budd [40] has compared the reliability of 3.2. Graph theory metrics ADS-B based information with that of proprietary A large set of graph theory metrics are available to airline information and classifies this data source as study network structures. The networks are modeled an “innovative and welcome source of empirical data”. as graphs with airport nodes and flight edges. The data Other studies have made use of the commercial OAG needed for this representation of the network structure database before for this purpose (e.g. [13, 37, 3, 29]). may come directly from airlines or aggregated flight The OAG database contains data on scheduled flights schedules. It is thus possible to collect the information and is known to be reliable. However, since OAG is at a very limited cost as it is generally available publicly. built around flight schedules, charter airline flights are We use eight network metrics to describe the network not included in the data provided by OAG [29]. We structures. These metrics have been chosen as they are opted to use the freely available ADS-B based flight well known and have previously been used to analyze information because it covers all operated flights, at a the structure of airline networks (e.g., [18, 41, 14]). low cost of data collection and because we are confident Four of these metrics describe the network as a whole that the data collected is accurate. and the four remaining metrics describe individual Airline business models and their network structures 5 nodes. The node-individual metrics are aggregated to to the node with the highest point centrality. The graph a network metric by calculating their network-specific centrality is then independent of the number of nodes Hirschman-Herfindal index (HHI). Thus, the analysis in the network. We deviate from this suggestion and focuses on the level differences between the node- calculate the HHI of the closeness node centralities individual metrics within a network. Furthermore, the in the network. Thereby, we put an emphasis on how HHI also normalizes the metric with regard to network equally distributed closeness centrality is. size. Node strength is the weighted sum of all incoming Average edge weight is the first network metric. It and outgoing edges in a node: [43]. is defined as the total number of flights over the total is the weight of the edge from i to j and is number of routes in the network. Hence, it expresses the set of neighbor nodes to i. The airport-specific how often on average the airline operates a flight on metric increases, as the airport is connected to more one of its routes. destinations and/or if existing routes are operated Latora and Marchiori [42] introduced network at higher frequencies. Very important airports such efficiency,ametricappliedfortheairlinenetworkcontext as FSC hubs will have a high node strength. Again, the variable analyzed per network is the HHI of node by Lin and Ban [43]. It is defined as , strength, hence the measure of how equally distributed the strength of the nodes in the network is. where N is the number of nodes in the network, and PageRank is an index to account for the importance is the distance between nodes i and j, measured of nodes in networks. PageRank is well-known today in the number of edges between them. If a directed as it is the basis for googles website scorings based network is fully connected, it will have on the work by Brin and Page [47]. It goes back to edges. In this case, the distance between all nodes is 1 the observation in social network analysis that the importance of a group leader not only results from and . the number of ties she has to other actors, but is also strongly influenced by the importance of her connected Edge density in a directed graph is defined as actors. With this observation, Katz [48] suggested an index to take both the number of edges as well as D , where M is the number of edges in the the neighboring nodes’ importance into account. In the airline context, PageRank is used to measure the graph. Edge density calculates the ratio of the existing likelihood that an arbitrary flight lands at a specific edges in a network over the maximum number of airport [49]. An airport is perceived important if this potential edges in the network. A complete graph has likelihood is high. It is intuitive that the likelihood that D = 1. Edge density has been used in the airline context a flight reaches an airport increases with the number among others by Sun and Wandelt [44]. of connected airports as well as the importance of the Finally, network transitivity is a metric to account connected airport. The individual nodes’ PageRank for clustering in a network. Formally, it is defined values are aggregated to a network metric by their HHI. Betweenness centrality is defined as the number as [45]. A triple of of shortest paths in the network that a node is part of. Following Freeman [46], the index calculation nodes refers to a node and its two directly connected starts by determining the number of existing shortest neighbors. The triangles on the graph refer to the paths from node i to node j in the network, denoted situation where three nodes are connected by a loop as . The number of these shortest paths going of three edges. Network transitivity accounts for how through node k is defined as . With these two often neighbors of one node are also connected directly. parameters, the betweenness centrality of node k It is a variant of a clustering coefficient. The first node-specific metric is closeness. It is reads . Betweenness centrality one of the well-known point centralities summarized in airline networks is commonly associated with the by Freeman [46]. It is calculated as . capability of an airport to offer connecting services (e.g. [16]). Freeman [46] suggests a graph betweenness That is, it is the inverse of the sum of all shortest centrality to account for the deviations from the distances from one node to all other nodes. This average node betweenness centralities. Again, we do metric is easily confounded with network efficiency, not make use of this index but calculate the HHI of yet the two metrics are far from identical. Important the node betweenness centrality to focus more on the to note is that the metric decreases as networks grow, inequality among nodes. hence comparing node closeness centrality across The metrics were selected such that they are networks needs to correct for network size. Freeman independent of the network size. The application of the [46] suggests measuring graph centrality based on the HHI eliminates the potential effect of the network size existing point centrality as the normalized difference in the data for the respective metrics. 6

It is important to keep in mind that none of the 2 reports the descriptive statistics of the calculated individual metrics will assume negative values. Table metrics.

Table 2: Descriptive statistics

Variable min median Mean max

Average edge weight 1.687 7.859 8.216 16.765

Network efficiency 0.048 0.124 0.153 0.395

Edge density 0.011 0.048 0.052 0.158

Transitivity 0.000 0.041 0.085 0.319

Closeness (HHI) 0.004 0.017 0.018 0.051

Node strength (HHI) 0.014 0.102 0.109 0.257

PageRank (HHI) 0.019 0.113 0.116 0.223

Betweenness (HHI) 0.028 0.209 0.306 1.000

3.3. Principal component analysis A principal component analysis (PCA) used as a This approach works particularly well if the original factorial method serves to reduce the dimensions in variables are sufficiently correlated. Handling and a dataset to a few principal components (PC). The interpreting the few linear combinations is easier than idea behind PCA is to find a few weighted linear dealing with the original variables. See Härdle and combinations of the multiple variables in the dataset Simar [50] for an introduction into the mathematical that nevertheless describe the dataset well; that is background of PCA. We follow Backhaus et al. [51] in they maximize the explained variance in the data. conducting this PCA.

Table 3: Correlation of network metrics Av P Be Ne Ed Tr No Cl ag ose tw an tw era de ge eR een si or ne de st ge an ti ke ren ss ns vi ne ed k( ty it (H ffi ss gt ge y HH ci h( HI (H we en I) HH ) HI cy ig ) ht I)

Averageedge weight 1-0.781 -0.502 -0.398 -0.220 0.490 0.485 0.423 Networkefficiency10.624 0.555 0.127 -0.638 -0.655 -0.568 Edgedensity 10.655 0.690 -0.393 -0.412 -0.323 Transitivity 10.129 -0.729 -0.746 -0.636 Closeness(HHI)10.180 0.165 0.185 Node strength (HHI) 10.973 0.953 PageRank (HHI) 10.895 Betweenness (HHI)1 Airline business models and their network structures 7

An essential question before conducting a PCA 3.5. Data analysis software is whether the multiple variables are sufficiently The analysis was conducted fully in the R software, correlated. Backhaus et al. [51] suggest using the version 3.5.1. The igraph package (version 1.2.1) for R Kaiser-Mayer-Olkin (KMO) criterion to test if the provides an extensive library of metrics for network matrix of correlation of the variables allows for a PCA. structure. PCA and cluster analysis are among the Generally, a PCA should not be conducted if the matrix statistical tools implemented in R. displays a KMO below 0.5. The weights assigned to the original variables to obtain the PC are referred to as factor loadings. The factor loadings support the interpretation of the 4. FINDINGS principal components. Factor loadings below a certain absolute value do not support the interpretation of the PC. Specifically, we shall not include absolute factor 4.1. Interpretation of factor loadings loadings below 0.25. Our data suggests this value as most variables load strongly to one of the PCs. Table 4: Factor loadings Dobruszkes [29] chooses an alternative approach by calculating a significance threshold based on the Principal component number of observations. It is not surprising from the description above 12 that some of the information aggregated in the eight Average edge weight 0.60 -0.40 metrics points towards the same aspects of network structure. Table 3 presents the correlation between Network efficiency -0.76 0.33 the different metrics and finds very strong effects for Edge density -0.47 0.79 some metrics such as node strength (HHI), PageRank (HHI) and betweenness (HHI). It is worthwhile to note Transitivity -0.79 (0.20) that closeness (HHI) and network efficiency hardly Closeness (HHI) (0.15) 0.93 correlate despite similarities in their definition at first sight. Our data shows a KMO of 0.69. Thus, we may Node strength (HHI)0.98(0.18) thus safely assume that the collected variables are PageRank (HHI)0.97 (0.15) sufficiently correlated to conduct a PCA. We turn to the Eigenvalue criterion to define the number of factors Betweenness (HHI)0.93(0.22) present in the data. Two PC have an Eigenvalue greater than 1, which indicates that they are suitable to explain In approaching an interpretation of the two factors the data. The following then proceeds with two PC. presented in Table 4, note that most metrics correlate Table 4 lists the PC loadings. Together, the two PC heavily with one factor. Network efficiency and explain 82% of the variance in the data (PC1 0.57 and transitivity load negatively, HHI of node strength, HHI PC2 0.24). of PageRank and HHI of betweenness centrality load positively to PC 1. Edge density and HHI of closeness influence PC 2 positively and PC 1 negatively. Average 3.4. Cluster analysis edge weight loads to both factors but with varying Given that we find two PC, we may conduct a cluster signs. It is the only metric negatively correlated with analysis in two dimensions, allowing for an intuitive PC 2, rendering it important for the discussion below. representation in a coordinate system. Cluster analysis PC 1 can be interpreted as the coverage ofthe network. is “one of the most used multivariate statistical An airline network will score high in this dimension techniques for segmentation” [52]. when node strength, PageRank and betweenness of The 58 airline network observations are positioned in the nodes in the network are unevenly distributed as a coordinate system expanded by PC 1 and PC 2. We are this will increase the respective HHI. Node strength interested in identifying groups of airlines with similar increases with many flights at an airport. If this airline structure. We thus conduct a 2-dimensional metric is unevenly distributed, this suggests that many cluster analysis based on a k-means algorithm. The airports have few flights and few airports operate many scree plot suggests four to eight clusters. Based on flights. This is achieved in a centrally oriented star-type the interpretability of the data, a solution with seven structure where few central airports serve a very large clusters is presented below. Since all eight original number of smaller airports. A betweenness HHI value metrics are defined in the positive range, a positioning equally suggests a star-type structure as the central of an airline in the negative range of PC 1 or PC 2 is airports lie on many shortest paths between other the result of negative factor loadings. airports while spoke airports do not have this property. A structure with some important airports would create a high value on the PageRank HHI. The PageRank of an airport increases in particular when it is connected 8 to few other important airports in the network. This is structure that differs between Clusters I and II and not achieved for instance, when several hub airports are the effect of the network size per se. Given that the interconnected. Furthermore, in order to score high on airlines spread further right on PC 1, this implies that the PC 1 dimension, network efficiency and transitivity the node strength, PageRank and betweenness are less should be low. Recall that network efficiency is the evenly distributed than in the networks of Cluster I. average of the inverse of the distances in the network. It indicates that the spread between the operations at The longer the paths in the network, the lower the the hubs in comparison to the served airports is more network efficiency values. Hence, airline networks pronounced than for Cluster I. It is noteworthy that where passengers can keep connecting from airport to Pegasus Airlines figures in Cluster I: It can rather be airport (even though it may not be very smart to do attributed to the LCC context, yet it operates a star- so for passengers) tend to have low network efficiency shaped network from a few Turkish airports. values. Transitivity is low when there are no clusters Closely related to these two is Cluster III that also of airports, i.e., neighboring nodes are not connected. includes FSCs. This cluster contains large airlines This is common for hub networks again. In a hub from smaller countries that act as their flag carrier, network, several spoke cities are connected to a hub, such as Ukraine International Airlines or Aer Lingus. but there typically is no direct connection between the The networks are operated at lower frequencies, that spokes. Summing up these points suggests that high is, lower edge weight, than the airlines in Cluster values on PC 1 relate to centrally oriented multi-hub I, moving Cluster III upwards on the PC 2 axis in structures where spoke airports are served distinctly comparison. Thus, based on the analysis of the network from one of the hubs. A network that scores high in metrics, the business model of FSC is separated into this dimension maximizes its coverage while limiting three subgroups. the number of routes. Cluster IV closely relates to Cluster V. Both clusters In contrast, PC 2 focuses on frequency of service. score similar values in PC 1 yet they differ in PC 2. Edge density and closeness HHI will contribute Cluster IV is positioned below Cluster V on the PC positively to PC 2. Edge density increases as the 2 axis. The networks of the airlines in Cluster IV are network moves from sparse to complete. Hence more more star-type than those of Cluster V. That is, the complete networks c.p. have a higher position on the Cluster IV networks rely more heavily on their hubs or PC 2 axis. Recall that node closeness is the inverse bases. The average flight frequency is greater. In sum, of the distance of a node to all others. To achieve a this suggests star-shaped networks with many airports high HHI few nodes need to acquire a major share of directly connected to a few important bases. Judging the total closeness centralities in the network. Highest from the business side, Cluster IV includes what is node closeness centrality is achieved by the center of a commonly perceived as large LCC on the European star and it decreases for multi-star networks. Average market: Ryanair, EasyJet, Lufthansa’s low cost edge weight furthermore negatively influences the offspring Eurowings, Air France’s offspring Hop! In positioning on the PC 2 axis. Average edge weight comparison, Cluster V contains mostly charter airlines, refers to the average number of flights per operated some of them with a regional focus. Their flights are route. Thus, airlines with a high frequency will score offered at a lower frequency. Keep in mind that the lower on PC 2 than ones with fewer flights on the collected data is from November in Europe, which is an routes. Bringing these three aspects together, a network off-peak season for holiday flights. During this season, scoring positively on the PC 2 axis offers many routes charter airlines will not operate at high frequencies on from one central base or hub but the routes are not few routes as it would be expected during the summer operated at high frequencies. months. Cluster VI is comparable to Cluster II on the PC 1 axis. It contains airlines with a rather regional focus 4.2. Findings from the cluster analysis of destinations. For instance, TAP Express and Iberia Figure 1 shows the positioning of the airlines in the Express both are regional feeder partners. Icelandair sample along the two axis. The seven resulting clusters focuses on flights from Iceland. However, Cluster VI are identified by colors and labelled by cluster number. airlines show a higher position on the PC 2 axis than Cluster I contains a subset of the traditional full service the airlines in Cluster II. The networks in this cluster carriers, including Air France, Iberia, British Airways have fewer destinations and flights are operated at and Lufthansa. The airlines share strong hub operations lower frequencies. and operate large networks. Cluster II features further Finally, Cluster VII contains three airlines whose full service carriers. This cluster subsumes the slightly networks have hardly any characteristics in common smaller counterparts to the airlines in Cluster I. These with the other clusters. Jet2.com is a low cost carrier airlines are positioned higher on PC 2 particularly operating with eight bases from within Great Britain. as their average edge weight is lower than their Loganair is a Scottish regional carrier with very counterparts in Cluster I. This is worth noting since particular services such as short connections to and the selected metrics are independent of network size. between Channel Islands. Rusline is a Russian regional Hence, what is observed here is a specific network carrier serving mostly Russian destinations. Airline business models and their network structures 9

Figure 1: Positioning of airlines (clusters are identified by color and number)

Observe that the far ends of the coordinate system flights. This difference in approach leads to different remain mainly void. Among the analyzed airline insights. Consider Ryanair and EasyJet in Figure 1. networks, only the airlines in Cluster VII score extreme Previously, Graham [53] has compared the network values on both components. The other airlines appear structures of LCC, including these two airlines. to emphasize only one component. Equally, Dobruszkes [29] includes these two airlines The airline industry is dynamic and the strategic in his sample on LCC. Both find relevant differences position of airlines in our sample has changed since the between Ryanair and EasyJet and it appears reasonable data collection. For instance, HOP! has been moving that differences between the two business models its operations significantly closer to Air France since appear more pronounced than for our study, which October 2018. Most notably, Germania Fluggesellschaft considers a broader sample of airlines to start with. mbH declared bankruptcy in February 2019 and ceased Graham [53] points out that at the time, Ryanair was operations the following day. Moreover, Flybe was sold serving more destinations than EasyJet, and at a to the Connect Airways consortium in February 2019 higher average frequency per route. At the same time, with the intention that Flybe will operate for Virgin Ryanair’s network relied strongly on its base in London Atlantic. Strategic changes such as these will affect Stansted while EasyJet distributed its flights almost the airlines’ networks and revisiting the networks evenly between London Stansted, Gatwick and Luton. in the future will most likely allow identifying the Dobruszkes [29] observes that Ryanair and EasyJet operational differences. differ in particular regarding his first two PC: the volume of supply and the density of the network. The difference in the other two PC, namely centralization 5. DISCUSSION and charter-like operations, is less significant. Our analysis takes into account not only LCC but reveals Our analysis intentionally rests solely on metrics of insights on the positions of more diverse airlines. Hence, network structure. Unlike, e.g. Dobruszkes [3, 29], it is natural that the differences between Ryanair’s and we do not include elements to describe the nature EasyJet’s networks appear less pronounced. In addition, of the airline network such as if the network serves we intended to correct our analysis for network size. ‘warm water’ destinations or the share of international Hence, the differences in network structure resulting 10 from the number of destinations served between the Europe. Further, we identified similarities to the results two competitors should be reduced in our analysis. of a global sample by Urban et al. [37]. It appears We do observe that EasyJet’s network scores slightly reasonable to assume that similar findings may be higher on the PC 1 dimension than Ryanair’s, which is generated for other geographic markets. partially driven by EasyJet operating not from one but from three key airports in the London area. The analysis of the data reveals that the network 6. CONCLUSIONS structures of FSC can be separated in three clusters. This is an indication that they do operate different We have collected and analyzed information on types of networks. When comparing solutions with European airline network structures. We did so with fewer clusters, we see that the next two clusters to the intention to discuss the benefit of detailed metrics be merged are Cluster II and III and subsequently for airline network structure for describing airline Cluster I will be added to this large cluster. Thus, it can business networks. We found that airlines of related easily be argued that the networks among the FSC are business models do share significant similarities in somewhat similar, yet there are three types of networks their network structures. In fact, clustering the airlines to be distinguished. The analysis identified large flag based on the information collected about their network carriers with more or less global networks with many structures by itself suggests a strikingly consistent spokes and potentially several hubs, hence positioned landscape of airline business models. Full service in lower ranges of PC 1. Furthermore, we find smaller carriers are found in Clusters I, II and partially III. Low counterparts to these used as feeders as well as smaller cost carriers form Cluster IV. Many charter airlines alliance partners. Network size is corrected for in the figure in Cluster V and most regional carriers in Cluster analyzed metrics, hence the finding here is that the VI. This observation supports the idea that studying smaller carriers share a similar structure. airline business networks can greatly benefit from the Urban et al. [37] conduct a cluster analysis only of objective assessment of airline network structure based FSC and LCC on a global scale. Their focus is on the on metrics from graph theory. convergence of these two business models. Parameters The airline network is commonly considered as one for the entire breadth of airline business models feed element of an airline business model. Jointly, all of the into their study. The route network is taken into account elements define an airline business model. Describing as a categorical variable (HS, PP, HS and PP, multi- the element of network more granularly is a step hub). Similar to our findings, they observe subgroups towards exploring the distinctions between business for both FSC and LCC. The identified groups for models. FSC are medium-size network carrier, global niche It is furthermore relevant to observe that the PCA market network carrier, high quality network carrier identifies two components that characterize airline and large-size network carrier. There is an overlap of networks. They hint at differences between HS und 13 airlines with our analysis. EasyJet, Ryanair, and PP networks. PC 1 aims at identifying the network Vueling are consistently clustered by our approach coverage whereas PC 2 points at the service network. jointly in Cluster IV such as in the group of PP LCC by Obviously, both aspects contribute to the understanding Urban et al. [37]. For the FSC, the allocation to clusters that a HS network increases its coverage and operates is not fully identical. Urban et al. [37] group Aeroflot, the connections between hubs at high frequency. Finnair, and TAP and among the global niche market From a methodological perspective, it is worth network carriers. They all show in Cluster II above with mentioning that PCA proves to be a powerful tool KLM joining them in our analysis. In contrast, Cluster to condense the network information obtained from I of large FSC includes Air France, British Airways, multiple metrics originating in graph theory. In general, Iberia, Lufthansa and Turkish Airlines while Urban the metrics are strongly correlated but not identical, as et al. [37] allocate them to three different subgroups is also the case with our data. This may be one reason of FSC, differentiating between medium- and large- why research has applied a wide variety of metrics to size and high quality. It is a relevant finding that both account for HS and PP networks. Trying to identify the approaches – covering the full breadth of airline one best metric among them is virtually impossible. business parameters and studying only the network Our analysis includes eight network metrics and the structure – produce mostly consistent findings. It PCA reduces them to two components. It is a valid would be interesting to extend our analysis similarly path for further research to validate that other network to a global sample to verify if subgroups among the metrics lead to consistent components. LCC can equally be identified. Omitting the regional The correlation of airline network structure and and charter airlines may also bring in additional focus business model is intuitive at first, given that it is on the LCC networks. common to assume that LCC operate PP networks As is true for all empirical studies, our findings only and FSC HS networks. However, it has been observed apply to the investigated dataset of European airlines. previously that these two types of networks cannot However, the airline business is mostly considered a meaningfully describe the existing spectrum of network global market and airline competition is not limited to structures. We have illustrated that incorporating more Airline business models and their network structures 11 granular and objective metrics of network structure in 12. Brueckner JK, Flores-Fillol R (2007) Airline the description of airline business models overcomes Schedule Competition. Rev Ind Organ 30 (3):161- this obstacle and allows to describe the network 177. doi:10.1007/s11151-007-9140-1 perspective better. 13. Burghouwt G, Hakfoort J, Ritsema van Eck J (2003) The spatial configuration of airline networks in Europe. Journal of Air Transport REFERENCES Management 9 (5):309-323. doi:10.1016/S0969- 6997(03)00039-5 1. Hu Y, Zhu D (2009) Empirical analysis of the 14. Guimerá R, Amaral LAN (2004) Modeling the worldwide maritime transportation network. world-wide airport network. The European Physica A: Statistical Mechanics and its Physical Journal B 38 (2):381-385. doi:10.1140/ Applications 388 (10):2061-2071. doi:https://doi. epjb/e2004-00131-0 org/10.1016/j.physa.2008.12.016 15. Guida M, Maria F (2007) Topology of the 2. Wang Y, Cullinane K (2016) Determinants of port Italian airport network: A scale-free small- centrality in maritime container transportation. world network with a fractal structure? Chaos, Transportation Research Part E: Logistics and Solitons & Fractals 31 (3):527-536. doi:10.1016/j. Transportation Review 95:326-340. doi:https:// chaos.2006.02.007 doi.org/10.1016/j.tre.2016.04.002 16. Wang J, Mo H, Wang F, Jin F (2011) Exploring the 3. Dobruszkes F (2006) An analysis of European network structure and nodal centrality of China’s low-cost airlines and their networks. Journal of air transport network: A complex network Transport Geography 14 (4):249-264. doi:https:// approach. Journal of Transport Geography 19 doi.org/10.1016/j.jtrangeo.2005.08.005 (4):712-721. doi:10.1016/j.jtrangeo.2010.08.012 4. Mishra S, Welch TF, Jha MK (2012) Performance 17. Alderighi M, Cento A, Nijkamp P, Rietveld indicators for public transit connectivity P (2007) Assessment of New Hub-and- in multimodal transportation networks. Spoke and Point-to-Point Airline Network Transportation Research Part A: Policy and Configurations. Transport Reviews 27 (5):529- Practice 46 (7):1066-1085. doi:https://doi. 549. doi:10.1080/01441640701322552 org/10.1016/j.tra.2012.04.006 18. Lordan O, Sallan JM, Simo P (2014) Study of the 5. Mattsson L-G, Jenelius E (2015) Vulnerability topology and robustness of airline route networks and resilience of transport systems – A discussion from the complex network approach: a survey and of recent research. Transportation Research Part research agenda. Journal of Transport Geography A: Policy and Practice 81:16-34. doi:https://doi. 37 (Supplement C):112-120. doi:10.1016/j. org/10.1016/j.tra.2015.06.002 jtrangeo.2014.04.015 6. Oum TH, Zhang A, Zhang Y (1995) Airline 19. Reggiani A, Signoretti S, Nijkamp P, Cento A Network Rivalry. The Canadian Journal of (2009) Network Measures in Civil Air Transport: Economics / Revue canadienne d’Economique A Case Study of Lufthansa. In: Naimzada AK, 28 (4a):836-857. doi:10.2307/135934 Stefani S, Torriero A (eds) Networks, Topology 7. Brueckner JK (2004) Network Structure and Dynamics: Theory and Applications to and Airline Scheduling. 52 (2):291-312. Economics and Social Systems. Springer Berlin doi:doi:10.1111/j.0022-1821.2004.00227.x Heidelberg, Berlin, Heidelberg, pp 257-282. 8. Alderighi M, Cento A, Nijkamp P, Rietveld P doi:10.1007/978-3-540-68409-1_14 (2005) Network competition – the coexistence 20. Bowen JT (2012) A spatial analysis of FedEx of hub-and-spoke and point-to-point systems. and UPS: hubs, spokes, and network structure. Journal of Air Transport Management 11 (5):328- Journal of Transport Geography 24 (Supplement 334. doi:10.1016/j.jairtraman.2005.07.006 C):419-431. doi:10.1016/j.jtrangeo.2012.04.017 9. Pels E (2008) Airline network competition: Full- 21. Button K, Ison S (2008) The economics of service airlines, low-cost airlines and long-haul low-cost airlines: Introduction. Research in markets. Research in Transportation Economics Transportation Economics 24 (1):1-4. doi:https:// 24 (1):68-74. doi:https://doi.org/10.1016/j. doi.org/10.1016/j.retrec.2009.01.008 retrec.2009.01.009 22. Daft J, Albers S (2015) An empirical analysis of 10. Babić D, Kalić M (2018) Modeling the airline business model convergence. Journal of selection of airline network structure in a Air Transport Management 46:3-11. doi:10.1016/j. competitive environment. Journal of Air jairtraman.2015.03.008 Transport Management 66:42-52. doi:https://doi. 23. Bachwich AR, Wittman MD (2017) The org/10.1016/j.jairtraman.2017.10.004 emergence and effects of the ultra-low cost 11. Brueckner JK (2010) Schedule Competition carrier (ULCC) business model in the U.S. airline Revisited. Journal of Transport Economics and industry. Journal of Air Transport Management Policy (JTEP) 44 (3):261-285 62:155-164. doi:10.1016/j.jairtraman.2017.03.012 12

24. Soyk C, Ringbeck J, Spinler S (2017) Long- 36. Pereira BA, Caetano M (2015) A conceptual haul low cost airlines: Characteristics of the business model framework applied to air business model and sustainability of its cost transport. Journal of Air Transport Management advantages. Transportation Research Part A: 44-45:70-76. doi:https://doi.org/10.1016/j. Policy and Practice 106:215-234. doi:https://doi. jairtraman.2015.02.006 org/10.1016/j.tra.2017.09.023 37. Urban M, Klemm M, Ploetner KO, Hornung M 25. Gillen D, Gados A (2008) Airlines within airlines: (2018) Airline categorisation by applying the Assessing the vulnerabilities of mixing business business model canvas and clustering algorithms. models. Research in Transportation Economics Journal of Air Transport Management. 24 (1):25-35. doi:http://dx.doi.org/10.1016/j. doi:10.1016/j.jairtraman.2018.04.005 retrec.2009.01.002 38. Dobruszkes F, Peeters D (2017) Where do 26. Wensveen JG, Leick R (2009) The long-haul low- planes fly past overhead? Determining cost carrier: A unique business model. Journal departure and arrival routes from radar traces. of Air Transport Management 15 (3):127-133. Applied Geography 89:173-183. doi:https://doi. doi:10.1016/j.jairtraman.2008.11.012 org/10.1016/j.apgeog.2017.10.009 27. Casadesus-Masanell R, Ricart JE (2010) From 39. Lange A, Sieling J, Gonzalez Parra G (2019) Strategy to Business Models and onto Tactics. Convergence in airline operations: The case Long Range Planning 43 (2):195-215. doi:https:// of ground times. Journal of Air Transport doi.org/10.1016/j.lrp.2010.01.004 Management 77:39-45. doi:https://doi. 28. Daft J, Albers S (2013) A conceptual framework org/10.1016/j.jairtraman.2019.03.001 formeasuringairlinebusinessmodelconvergence. 40. Budd LCS (2012) The potential use of consumer Journal of Air Transport Management 28:47-54. flight radar apps as a source of empirical data: A doi:10.1016/j.jairtraman.2012.12.010 study of A320 utilization rates in Europe. Journal 29. Dobruszkes F (2013) The geography of European ofAirTransportManagement25:53-56.doi:https:// low-cost airline networks: a contemporary doi.org/10.1016/j.jairtraman.2012.08.004 analysis. Journal of Transport Geography 41. Dunn S, Wilkinson SM (2016) Increasing 28:75-88. doi:https://doi.org/10.1016/j. the resilience of air traffic networks using a jtrangeo.2012.10.012 network graph theory approach. Transportation 30. Klophaus R, Conrady R, Fichert F (2012) Low Research Part E: Logistics and Transportation cost carriers going hybrid: Evidence from Europe. Review 90:39-50. doi:https://doi.org/10.1016/j. Journal of Air Transport Management 23:54-58. tre.2015.09.011 doi:10.1016/j.jairtraman.2012.01.015 42. Latora V, Marchiori M (2001) Efficient 31. Lohmann G, Koo TTR (2013) The airline Behavior of Small-World Networks. Physical business model spectrum. Journal of Air Review Letters 87 (19):198701. doi:10.1103/ Transport Management 31:7-9. doi:10.1016/j. PhysRevLett.87.198701 jairtraman.2012.10.005 43. LinJ, BanY(2014) Theevolvingnetworkstructure 32. Jean DA, Lohmann G (2016) Revisiting the of US airline system during 1990–2010. Physica airline business model spectrum: The influence of A: Statistical Mechanics and its Applications post global financial crisis and airline mergers in 410:302-312. doi:10.1016/j.physa.2014.05.040 the US (2011−2013). Research in Transportation 44. Sun X, Wandelt S (2014) Network similarity Business & Management 21:76-83. doi:10.1016/j. analysis of air navigation route systems. rtbm.2016.06.002 Transportation Research Part E: Logistics and 33. Moir L, Lohmann G (2018) A quantitative Transportation Review 70:416-434. doi:10.1016/j. means of comparing competitive advantage tre.2014.08.005 among airlines with heterogeneous business 45. Girvan M, Newman MEJ (2002) Community models: Analysis of U.S. airlines. Journal of Air structure in social and biological networks. Transport Management 69:72-82. doi:https://doi. Proceedings of the National Academy of Sciences org/10.1016/j.jairtraman.2018.01.003 99 (12):7821-7826. doi:10.1073/pnas.122653799 34. Mason KJ, Morrison WG (2008) Towards a 46. Freeman LC (1978) Centrality in social networks means of consistently comparing airline business conceptual clarification. Social Networks 1 models with an application to the ‘low cost’ airline (3):215-239. doi:10.1016/0378-8733(78)90021-7 sector. Research in Transportation Economics 24 47. Brin S, Page L (1998) The anatomy of a large- (1):75-84. doi:10.1016/j.retrec.2009.01.006 scale hypertextual Web search engine. Computer 35. O’Connell JF (2011) The rise of the Arabian Networks and ISDN Systems 30 (1):107-117. Gulf carriers: An insight into the business model doi:10.1016/S0169-7552(98)00110-X of Emirates Airline. Journal of Air Transport 48. Katz L (1953) A new status index derived from Management 17 (6):339-346. doi:https://doi. sociometric analysis. Psychometrika 18 (1):39-43. org/10.1016/j.jairtraman.2011.02.003 doi:10.1007/bf02289026 Airline business models and their network structures 13

49. Conti E, Cao S, Thomas AJ (2013) Disruptions in 52. Wen C-H, Lai S-C (2010) Latent class models of the U.S. Airport Network. arXiv international air carrier choice. Transportation 50. Härdle W, Simar L (2007) Applied Multivariate Research Part E: Logistics and Transportation Statistical Analysis. Springer Berlin Heidelberg, Review 46 (2):211-221. doi:https://doi. Berlin, Heidelberg. doi:10.1007/978-3-540- org/10.1016/j.tre.2009.08.004 72244-1_9 53. Graham M (2009) Different models in different 51. Backhaus K, Erichson B, Plink W, Weiber R spaces or liberalized optimizations? Competitive (2016) Multivariate Analysemethoden, vol 14. strategies among low-cost carriers. Journal of Springer Gabler, Berlin Heidelberg Transport Geography 17 (4):306-316. doi:https:// doi.org/10.1016/j.jtrangeo.2009.02.004

APPENDIX:

Airline Code Country Number of aircraft in dataset Aegean Airlines A3 Greece 47 Aer Lingus EI Ireland 63 Aeroflot SU Russia 224 Air Baltic BT Latvia 30 Air Europa UX Spain 52 Air France AF France 230 Air Serbia AF France 230 Alitalia AZ Italy 120 Atlasjet KK Turkey 24 Austrian Airlines OS Austria 84 Azur Air ZF Russia 22 Belavia B2 Belarus 29 Blue Air 0B Romania 30 British Airways BA United Kingdom 318 Brussels Airlines SN Belgium 51 Condor DE Germany 42 Eastern Airways T3 United Kingdom 30 EasyJet U2 United Kingdom 280 Eurowings & Germanwings* EW/U2 Germany 78 Finnair AY Finland 64 Flybe BE United Kingdom 84 Germania ST Germany 29 HOP! A5 France 86 Iberia IB Spain 114

Iberia Express I2 Spain 21 Icelandair FI Iceland 32 14

Jet2 LS United Kingdom 75 KLM KL The Netherlands 160 Loganair LM United Kingdom 28 LOT LO Poland 47 Lufthansa LH Germany 280 Lufthansa Cityline LH Germany 50 Norwegian DY Norway 142 Pegasus Airlines PC Turkey 74 Rossiya FV Russia 62 Rusline 7R Russia 23 Ryanair FR Ireland 417 S7 Airlines S7 Russia 78 SAS SK Sweden 162 SunExpress XQ Turkey 46 Swiftair WT Spain 38 Swiss LX Switzerland 77 TAP Express NI Portugal 22 TAP Portugal TP Portugal 66 Tarom RO Romania 25 Thomas Cook Airlines MT United Kingdom 60 Transavia HV The Netherlands 70 TUI fly HV The Netherlands 70 Turkish Airlines TK Turkey 324 Ukraine Int. Airlines PS Ukraine 41 Ural Airlines U6 Russia 43 UTair UT Russia 68 Virgin Atlantic Airways VS United Kingdom 38 Volotea V7 Spain 25 Vueling VY Spain 107 Wideroe WF Norway 41 Wizz Air W6 Hungary 87 Ya mal Airlines YC Russia 40

Table A.1: List of airlines, *operated jointly at the time of data collection